Thinning out Steiner Trees

Matteo Fischetti

Markus Leitner

Ivana Ljubic

Martin Luipersbeck

Michele Monaci

Max Resch

Domenico Salvagnin

Markus Sinnl

University of Padova and Vienna

Aussois, January 2015 1

• The 11th DIMACS challenge was on Steiner Tree Problems

Aussois, January 2015 2

Vienna and Padua together

Aussois, January 2015 3

• A bunch of effective codes initially provided by Vienna’s team• All codes re-engineered and re-tuned to improve performance• New exact and heuristic codes implemented (this talk)• Initial filter to select the actual code to be run based on the instance type

• Four codes (all based on Cplex 12.6) finally submitted to DIMACS’s challenge

Challenge results

Aussois, January 2015 4

Detailed final results posted by Gerald Gamrath at

http://dimacs11.cs.princeton.edu/contest/challenge-results.pdf

Many variants / scores not a single winner…

… however #MozartBalls ranked first in many categories

Steiner Tree Problem

Aussois, January 2015 5

Hypercube instances• Since the beginning of our study we focused on the (in)famous

hypercube instances introduced by

Isabel Rosseti, Marcus Poggi de Aragao, Celso C. Ribeiro, Eduardo Uchoa, and Renato F. Werneck. New

benchmark instances for the Steiner problem in graphs. In Extended Abstracts of the 4th Metaheuristics

International Conference (MIC’2001), pages 557–561, Porto, 2001.

• Real terminals only• All edges have cost 1• Very symmetrical

very difficult LPs

very hard even for heuristicsAussois, January 2015 6

Looking for more effective models

• Many hard instances (including hc*u) involve uniform edge costs

• For these instances, edge costs can easily be moved to nodes …

… hence edge variables become redundant (actually, harmful as they add a lot of symmetry and overload the model with useless cuts)

• We better work on the space of node-variables only …

… and only impose connectivity of the subgraph induced by the selected nodes, as in

Eduardo Alvarez-Miranda, Ivana Ljubic, and Petra Mutzel. The maximum weight connected subgraph problem.

In Facets of Combinatorial Optimization: Festschrift for Martin Groetschel, 245–270, Springer, 2013.

Aussois, January 2015 7

Node separators

Aussois, January 2015 8

Model for uniform edge costs

Aussois, January 2015 9

Connectivity cut separation

Aussois, January 2015 10

Basic model

Aussois, January 2015 11

Branch & Cut code

• Built on top of IBM ILOG Cplex 12.6

• Initial compact relaxation basic model (quite tight for some instances)

• Very basic implementation with connectivity cuts separated for integer points only (lazycut callback)

• Cplex’s cuts at default level (but 0-1/2 cuts that are set to aggressive)

• Cplex’s primal heuristics at default level (local-branching set to on)

• More elaborated implementations tried worse performance

• Local branching heuristic using the B&C itself as a black-box solver

– invoked a few times before attempting the exact solution

– used to gather primal information (better and better feasible solutions) as well as dual information (connectivity cuts) restart policy

Aussois, January 2015 12

Comparing exact models

Aussois, January 2015 14

Benders-like heuristic

Aussois, January 2015 15

Set covering is everywhere!• For pure instances (real terminals only) the node model can be interpreted

as a huge set covering problem

• Given the basic model + pool of connectivity cuts, the relaxation in the Benders-like framework can then be solved by a specialized set-covering heuristic we used the CFT heuristic of

Alberto Caprara, Matteo Fischetti, and Paolo Toth. A heuristic method for the set covering

problem. Operations Research, 47(5):730–743, 1999.

• For non-uniform instances, blurred version in the “relaxation”, edge costs are heuristically split among adjacent nodes (so the node model applies)

• Very fast and effective, in particular, for hc*u instancesAussois, January 2015 17

Set-covering heuristic results

Aussois, January 2015 18

Thanks for your attention

• Full paper

M. Fischetti, M. Leitner, I. Ljubic, M. Luipersbeck, M. Monaci, M. Resch, D. Salvagnin, and M. Sinnl, Thinning out Steiner trees: a node-based model for uniform edge costs, Technical Report DEI, University of Padua, 2014.

and slides available at

http://www.dei.unipd.it/~fisch/papers/

http://www.dei.unipd.it/~fisch/papers/slides/

Aussois, January 2015 19